thanks Dan for your codes. i saw it in action and helped me to understand
the commands. following your posting, i have also had the opportunity to run
the routine in SAS elsewhere. when i saw the output i realised that:
/*
In the next step we create a special dataset that contains the 44 calculated
variances of logit sensitivity and the 44 variances of specificity in each
study. Three additional records are created which contain the starting
values for the 3 additional variance parameters of the bivariate model:
record 1 holds the starting value for the between-study variance in logit
sensitivity, record 2 the covariance between logit sensitivity and
specificity, and record 3 the between-study variance in logit specificity
*/
data cov;
/* start the file with three starting values for (co)variance parameters of
the random effects (zero works well) SAS expects the name est for this
variable
*/
if _n_ eq 1 then do;
est = 0; output; est =0; output; est = 0; output;
end;
/* followed by the 88 calculated variances of sensitivity and specificity of
each study */
set bi_meta;
est = var_logit; output;
keep est;
run;
does not concatenate the 3 new observations (appended) to the bi-meta.dta
instead it creates a separate covariance data set i.e. cov.dta with 91
variables, where the first three observations are three trailing zero's (see
below output from SAS) before the 88 variances.
NOTE: There were 88 observations read from the data set WORK.BI_META.
NOTE: The data set WORK.COV has 91 observations and 1 variables.
NOTE: DATA statement used (Total process time):
real time 0.01 seconds
cpu time 0.01 seconds
having come this far and the data sets (bi_meta.dta and cov.dta) are
prepared (i think! and many thanks to Dan Blanchette's help) for gllamm, i
need to ask those in the know about gllamm (Stata) and Proc Mixed (SAS) as
to what would be the equivalent of the following SAS's Proc Mixed in Stata's
gllamm (if any):
1. settings of degrees of freedom to obtain p-values based on normal
distribution rather than t-distribution,
2. random effects with possible correlation (i.e. unstructured covariance
structure),
3. 'repeat' statement,
4. parms (i.e. covariance parameters) i.e. making use of the cov.dta file
(and how in Stata?)
i have included again the Proc Mixed statement that was appendixed on the
paper i quoted earlier in the thread.
proc mixed data=bi_meta method=reml cl ; /* option cl will give confidence
intervals */
/* study_id and modality are categorical variables */
class study_id modality;
/* model statement: asking for different estimates of mean sensitivity and
specificity for each modality, provide large value for degrees of freedom to
obtain p-values based on normal distribution rather than the t-distribution
(=default in SAS)
*/
my understanding (again, i think!) so far is that SAS's Proc Mixed 'class'
is handled by Stata's gllamm i(varlists) whilst interaction terms
dis*modality and non_dis*modality are generated before calling gllamm. the
noint option in SAS Proc Mixed would be nocons in Stata's gllamm while 95%
CI is an a default setting in Stata's gllamm. however i cannot built a
gllamm model that give an output as published in the article:
Imaging modality Mean sensitivity (95% CI) Mean specificity (95% CI) Mean
DOR (95% CI)
LAG 0.67 (0.57 to 0.76) 0.80 (0.73 to 0.85) 8.13 (5.16 to 12.82)
CT 0.49 (0.37 to 0.61) 0.92 (0.88 to 0.95) 11.34 (6.66 to 19.30)
MRI 0.56 (0.41 to 0.70) 0.94 (0.90 to 0.97) 21.42 (10.81 to 42.45)
Imaging modality Mean sensitivity (95% CI) Mean specificity (95% CI) Mean
DOR (95% CI)
LAG 0.67 (0.57 to 0.76) 0.80 (0.73 to 0.85) 8.13 (5.16 to 12.82)
CT 0.49 (0.37 to 0.61) 0.92 (0.88 to 0.95) 11.34 (6.66 to 19.30)
MRI 0.56 (0.41 to 0.70) 0.94 (0.90 to 0.97) 21.42 (10.81 to 42.45)
P-value LAG vs. CT 0.023 0.0002 0.35
P-value LAG vs. MRI 0.23 0.0001 0.021
P-value CT vs. MRI 0.47 0.34 0.15
i've looked at www.gllamm.org website to check gllamm's manual (pdf) but
unless i am missing something (e.g. very real lack of understanding of the
model) i could not find a way to apply gllamm in this situation. again, i am
grateful for advice/help on how to proceed from here on and also, needless
to say, apology for an extensive and perhaps over-verbose postings.
